1
On Buffer Sizing for Voice in 802.11 WLANs
David Malone, Peter Clifford, Douglas J. Leith
Hamilton Institute, NUI Maynooth
70
60
50
Throughput (kbps)
Abstract— The use of 802.11 to transport delay sensitive traffic
is becoming increasingly common. This raises the question of the
tradeoff between buffering delay and loss in 802.11 networks. We
find that there exists a sharp transition from the low-loss, lowdelay regime to high-loss, high-delay operation. Given modest
buffering at the access point, this transition determines the voice
capacity of a WLAN and its location is largely insensitive to the
buffer size used.
40
30
sta, buf=1
sta, buf=2
sta, buf=5
sta, buf=10
AP, buf=1
AP, buf=2
AP, buf=5
AP, buf=10
20
I. I NTRODUCTION
10
IEEE 802.11 technology has been enormously successful,
with wireless 802.11a/b/g edge networks now being very
common. While data traffic (web, email, media downloads
etc) currently constitutes the bulk of Internet traffic, voice
applications are becoming increasingly important. Voice traffic
differs fundamentally from data traffic in its sensitivity to
delay and loss. This has led to substantial interest in ensuring
appropriate quality of service (QoS) for voice traffic in mixed
voice and data networks, including the development of the
recent 802.11e standard specifically targeted at addressing
QoS issues. However, the focus of published work has been
largely on MAC design and operation to ensure appropriate
prioritisation of delay-sensitive traffic. To our knowledge,
almost no published work exists on the question of appropriate
network buffer sizing for voice traffic in 802.11 WLANs.
In this paper we investigate buffer sizing for voice calls in
802.11 networks. Of course, there have been many simulation
and modelling studies of 802.11 networks. While some of
these studies have considered voice traffic (e.g. [1], [2], [3]),
including some commenting upon the value of queueing voice
separately from other traffic (e.g. [4]), to our knowledge the
present paper is the first to address the question of network
buffer sizing for voice traffic. At the application layer, playout
buffering has been considered for 802.11 (e.g. [5]), but this
is a separate issue from network layer buffer sizing. In [6] it
is observed that increased buffer sizing does not necessarily
improve the performance of inelastic traffic.
0
II. B UFFER
SIZING FOR VOICE
We consider an infrastructure mode WLAN where traffic is
routed via an access point (AP). Following [7], we model a
two-way voice call as a 64kbs on-off traffic stream with on and
off periods exponentially distributed with mean 1.5s, subject
to a minimum of 240ms. Traffic passes between between a
wireless client station and a device behind the AP. To account
for the two-way correlated nature of voice conversations; the
on/off periods of one half of a call correspond to the off/on
periods of the other. We consider an 802.11b PHY with the
Work supported by Science Foundation Ireland grant IN3/03/I346.
0
5
10
15
20
25
30
Number of conversations
Fig. 1. Achieved throughput for AP/client voice with various buffer sizes as
the no. of calls is increased. ns simulation.
following MAC parameters: 20 µs slot time, CWmin 32, DIFS
50 µs, SIFS 10 µs, long 192 µs preamble, 100 byte packets
(including headers) simulated with ns [8].
Figure 1 shows the average throughput per call as we
increase the number of voice conversations (and so stations)
in the network. Values are shown both for the aggregate of
client stations and the AP and results are given for buffer
sizes of 1, 2, 5 and 10 packets with the buffer in the AP
set to be the same size as in each of the stations. We can
see immediately that the throughput achieved by the AP falls
relative to that of the aggregate client stations as the number
of calls is increased. This is perhaps unsurprising as the
802.11 MAC enforces per station fairness; that is, the client
stations and the AP each win approximately the same number
of transmission opportunities despite the fact that the AP
carries n times as much traffic as each client station. The
situation with on-off traffic is complicated by the fact that,
firstly, voice traffic is relatively low rate and so need not
make use of every available transmission opportunity awarded
by the 802.11 MAC. Secondly, a voice conversation involves
speakers approximately taking turns at talking. That is, traffic
is between pairs of speakers with the on period of one speaker
roughly corresponding to the off period of the other. Both of
these features mitigate the contention between the wireless
stations and the AP for access to the wireless channel. Hence,
while a simple argument based on per station fairness would
suggest that the AP throughput would be 1/(n + 1) that of
the aggregate client stations, it can be seen from Figure 1
that this is not the case1 . This observation is not new and has
been discussed elsewhere[2]. Previous work has not, however,
considered the impact of buffer sizing on network behaviour.
1 We comment that AP throughput does scale as 1/(n+1) that of the client
stations when the WLAN nodes are saturated, see for example [2].
2
AP
1
0.9
0.8
0.8
0.7
0.7
0.6
0.6
0.5
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0
0.002
0.004
0.006
0.008
0.01
0.012 0.014
packet inter−arrival time (s)
0.016
0.018
0
0.02
60
50
0.5
0.4
0
70
0
0.002
0.004
0.006
0.008
0.01
0.012 0.014
packet inter−arrival time (s)
0.016
0.018
0.02
Throughput (kbps)
CDF
CDF
client station
1
0.9
40
30
sta, buf=1
sta, buf=2
sta, buf=5
sta, buf=10
AP, buf=n
AP, buf=2n
AP, buf=5n
AP, buf=10n
20
Fig. 2.
Client and AP packet inter-arrival time cumulative distribution
functions for the number of voice calls, n = 10.
10
0
0
5
10
15
20
25
30
Number of conversations
AP
0.2
0.18
Fig. 4. Achieved throughput for AP/client voice with various buffer sizes as
the number of conversations is increased. AP buffer scaled with number of
calls.
0.16
0.14
frequency
0.12
0.1
0.08
0.06
0.04
0.02
0
0
1
2
3
4
5
#arrivals
6
7
8
9
10
Fig. 3. Distribution of number of arrivals in a 10ms interval at an AP with
10 on/off calls in progress.
We can see from Figure 1 that the choice of buffer size
has a strong impact on the throughput achieved by the AP.
For very small buffers it can be seen that the AP throughput
falls to around 90% of the 64Kbs offered load from a voice
call by the time two calls are active. This drop in throughput
is already likely to yield unacceptable quality of service; that
is, to restrict the network voice-call capacity to two calls or
less. A buffer length of 10 packets improves AP throughput
significantly up to about 10 conversations, thereby greatly
increasing the network voice-call capacity compared to the
situation when very small buffers are used.
We can gain insight into this behaviour by considering
the arrival processes at the client stations and AP in more
detail. The arrival process at a client station consists of on-off
64Kbs traffic. During an on-period packets arrive at regular
10ms intervals; no packets are generated during an off-period.
The measured cumulative distribution function of packet interarrival times is shown in Figure 2. Since the inter-arrival times
are always greater than or equal to 10ms, stability is assured
provided the mean service time at the network interface queue
of a client station is less than 10ms. In contrast, the arrival
process at the AP is the aggregate of n on-off arrival processes,
corresponding to the n voice call halves. It may happen that
packets from several calls arrive at the AP within a short
time of each other and thus the inter-arrival times at the AP
queue are not lower-bounded by 10ms. This is evident in the
measured cumulative distribution of packet inter-arrival times
shown in Figure 2. As a result, the queue sizing requirement
at the AP differs from that at the clients.
We note, however, that since the inter-arrival times for each
individual call are at least 10ms, the number of packets that
can arrival at the AP during a 10ms interval is no more than
n. This is a worst case bound and may occur only rarely. For
example, Figure 3 shows the measured distribution of packet
arrivals at the AP in a 10ms interval. This simple analysis
suggests that the AP buffer size should be set equal to at least
the number of calls n.
The impact of this change is demonstrated in Figure 4 where
we scale the AP’s buffer to be n times the size of the stations’
buffers. Figure 5 shows the corresponding delay. A number of
observations can be seen immediately:
1) The per call throughput is close to 64Kbs and the mean
delay is below 10ms for up to 12 calls. For greater
than 12 calls, the per call throughput falls rapidly —
by 13 calls the throughput has fallen by more than 10%
— and delay quickly rises. This is likely to yield an
unacceptable call quality i.e. the voice-call capacity of
the network is therefore approximately 12 calls. This is
in good agreement with a back-of-envelope calculation
based on [9] which indicates a voice capacity upper
limit (neglecting packet collisions and other contention
overhead) of around 15 calls.
2) There is an abrupt transition from the low-loss, lowdelay regime to high-loss, high-delay operation. Below this transition, buffer sizing has little impact on
throughput and delay for up to 12 calls. Above 12 calls,
throughput falls below 90% of offered load for all sizes
of buffer and delay rises rapidly. That is, the location of
the transition is essentially independent of the level of
buffer provisioning and thus network capacity is fixed
at approximately 12 calls regardless of buffer size.
3) In the high-loss, high-delay regime above 12 calls, the
total delay depends strongly on buffer size. This is to be
expected as in this unstable regime the buffer contains
a standing queue that scales with buffer size.
4) Smaller buffer sizes yield shorter delays in the lowthroughput, high-delay regime with greater than 12 calls.
We note that across a wide range of situations including
peer-to-peer networks, infrastructure mode networks, plain
3
0.14
sta, buf=1
sta, buf=2
sta, buf=5
sta, buf=10
AP, buf=n
AP, buf=2n
AP, buf=5n
AP, buf=10n
0.12
Total Delay (s)
0.1
0.08
0.06
0.04
0.02
0
0
5
10
15
20
25
30
Number of conversations
Fig. 5.
Total delay (queueing delay plus MAC delay) as the no. of
conversations is increased. AP buffer scaled with no. of calls.
0.016
sta, buf=1
sta, buf=2
sta, buf=5
sta, buf=10
AP, buf=n
AP, buf=2n
AP, buf=5n
AP, buf=10n
Interpacket time
0.014
MAC Delay (s)
0.012
0.01
0.008
0.006
the wireless channel is occupied by transmissions (i.e., the
time the countdown is halted because the channel is sensed
busy). It is only through these quantities that the buffer size
can have an impact on MAC delay.
With on-off traffic such as voice, it can readily be seen
that when a station’s queue is not backlogged some transmission opportunities are not used because there is no packet
available to send. However, if the queue becomes backlogged,
then the number of unused transmission opportunities must
decrease. Consequently, we expect both the frequency of
packet collisions and the time that the channel is occupied by
transmissions to increase. Thus, the inter-packet service time
(MAC delay) of the interface queues actually increase as the
queues becomes backlogged. Conversely, the queue tends to
backlog as the inter-packet service time increases. Therefore
the potential exists for a reinforcing feedback whereby the onset of queueing leads to further queue buildup and instability.
We note that the buffer sizing problem in 802.11 WLANs is
fundamentally different from its wired analogue because of
this complex feedback loop which couples service rate and
queueing.
III. C ONCLUSIONS
0.004
0.002
0
0
5
10
15
20
Number of conversations
25
30
Fig. 6.
MAC level delay with various buffer sizes as the number of
conversations is increased.
802.11, prioritised 802.11e, pure voice environments and
mixed voice/data environments, we have observed broadly
similar behaviour and find that there consistently exists a
sharp transition from low-loss, low-delay operation to a highloss, high-delay regime. This transition determines the voice
capacity of the network and its insensitivity to buffer sizing is
surprising.
We can gain more insight into this behaviour by looking
at the MAC component of the delay, shown in Figure 6.
The MAC delay is the mean time spent by the MAC layer
transmitting a packet. This includes collisions, contention,
transmission and acknowledgement. Thus, the MAC delay is
the inter-packet service time of a station’s interface queue. The
MAC delays are naturally much shorter than the total delays,
though their full impact will be scaled by the buffer size. In
Figure 6 we can clearly see that increasing buffer sizes results
in an increased MAC delay, particularly for loaded networks2.
The MAC delay is determined by the following quantities:
(i) the 802.11 contention window value, which doubles each
time a packet transmission fails due to a collision, (ii) the
number of transmission failures due to collisions that occur
before a transmission succeeds, and (iii) the duration for which
2 We can also see that once congested the MAC delay increases approximately linearly as additional stations are added. This is because the 802.11
MAC mechanism is distributing the available transmission time approximately
evenly among all stations.
We have considered the tradeoff between buffering and loss
for voice traffic in 802.11 networks. We find that there exists a
sharp transition from the low-loss, low-delay regime to highloss, high-delay operation. This transition determines the voice
capacity of a WLAN and its location is largely insensitive to
the buffer size used. Interestingly, this observation indicates
that recently proposed finite-load analytic models for 802.11
networks with small buffers [1] can be employed to accurately
predict network capacity even when large buffers are used. We
have identified a complex feedback loop between service rate
and queueing in 802.11 that contributes to this sharp transition.
R EFERENCES
[1] K. Duffy, D. Malone, and D. Leith, “Modeling the 802.11 Distributed Coordination Function in non-saturated conditions,” IEEE Communications
Letters, vol. 9, no. 8, pp. 715–717, 2005.
[2] P. Clifford, K. Duffy, D. Leith, and D. Malone, “On improving voice
capacity in 802.11 infrastructure networks,” in Proc. WIRELESSCOM,
2005.
[3] N. Hegde, A. Proutiere, and J. Roberts, “Evaluating the voice capacity
of 802.11 WLAN under distributed control,” in Proc. IEEE LANMAN,
September 2005.
[4] J. Yu, S. Choi, and J. Lee, “Enhancement of VoIP over IEEE 802.11
WLAN via dual queue strategy,” in Proc. IEEE ICC, 2004.
[5] P. Bellavista, A. Corradi, and C. Giannelli, “Adaptive buffering-based
on handoff prediction for wireless internet continuous services,” in Proc.
HPCC, 2005.
[6] S. Selvakennedy, “The impact of transmit buffer on EDCF with framebursting option for wireless networks,” in Proc. IEEE Local Computer
Networks, 2004.
[7] A. Markopoulou, F. Tobagi, and M. Karam, “Assessing the quality of
voice communications over internet backbones,” IEEE Transactions on
Networking, vol. 11, no. 5, pp. 747–760, October 2003.
[8] S. Wiethölter and C. Hoene, “Design and verification of an IEEE 802.11e
EDCF simulation model in ns-2.26,” Technische Universität Berlin, Tech.
Rep. TKN-03-019, November 2003.
[9] D. Hole and F. Tobagi, “Capacity of an IEEE 802.11b wireless LAN
supporting VoIP,” in International Conference on Communications, 2004.